Pixelesq

Pixelesq Agent Skills

[smithery badge]

Manage websites built on Pixelesq from any AI agent. Build and edit pages section by section, write content into collections, tune SEO and structured data, read Search Console and site analytics, manage redirects and images, restyle the theme, and publish when the user says so.

This repository is three things at once:

• Seven skills in the Agent Skills format, loaded by Claude Code, Cursor, Codex, Copilot, Gemini CLI, Goose and other compatible agents.
• Four specialist agents for Claude Code (and any platform that reads agents/*.md).
• The Pixelesq MCP connector, declared in .mcp.json: https://mcp.pixelesq.app/mcp, 70 tools, OAuth sign-in with your Pixelesq account.

Install

Claude Code (plugin)

/plugin marketplace add pixelesq/pixelesq-skills/plugin install pixelesq@pixelesq

Then run /mcp and sign in to Pixelesq. The skills load on demand, the agents are available as @pixelesq:site-manager, @pixelesq:seo-specialist, @pixelesq:content-strategist and @pixelesq:design-consultant, and the connector is registered from .mcp.json.

Claude Code (connector only)

claude mcp add --transport http pixelesq https://mcp.pixelesq.app/mcp

Claude.ai and Claude Desktop

Settings, Connectors, Add custom connector, paste https://mcp.pixelesq.app/mcp, sign in.

ChatGPT

Open Pixelesq in the ChatGPT plugin directory and connect; that listing uses https://pixelesq-mcp.pixelesq.workers.dev/mcp, the same server on its older address. To add the server by hand: Settings, turn on Developer mode (under Apps or Security and login), then Plugins, +, Add custom MCP server, paste https://mcp.pixelesq.app/mcp, set authentication to OAuth and install it. A connection made on one address does not carry over to the other.

Cursor

Add the server to mcp.json:

{ "mcpServers": { "pixelesq": { "url": "https://mcp.pixelesq.app/mcp" } } }

VS Code

{ "servers": { "pixelesq": { "type": "http", "url": "https://mcp.pixelesq.app/mcp" } } }

Other platforms

Clone the repository, point the agent's skills directory at skills/, and configure the MCP server URL https://mcp.pixelesq.app/mcp. Clients register dynamically or present a client ID metadata document; both are accepted.

Sign-in and scopes

The connector uses OAuth 2.1 with PKCE. Sign in with Google or with a Pixelesq email and the code it sends. A Pixelesq account with at least one project is required; sign up at https://app.pixelesq.ai.

Scopes are granted at sign-in and stamped into the token: projects:read, content:read, content:write, publish, analytics:read. A tool outside the grant answers with a refusal that asks the client to reconnect.

How edits stay safe

• Saves to pages, sections, headers, footers and entries change a draft, and publish_page, publish_partial and publish_entry make it live. A few writes change the live site right away: update_page_meta on a published page, the redirect tools, update_collection_schema, update_entry (a new name shows in listings), delete_page, delete_entry, update_theme (unless publish: false) and update_form_schema (a form has no draft). The skills tell the agent to confirm before any of these.
• Every save carries the lastVersion from the read before it; a stale version is refused instead of overwriting someone else's work.
• A publish carries expectedRevisionId and expectedHeadAnchor from the read the user reviewed, and refuses if the draft moved.
• Saves merge field by field, arrays replace whole, null clears a field. The agent only touches what the user asked for.
• Structural section edits (insert, remove, reorder, bind) take an idempotencyKey so a lost answer can be retried safely.

Skills

Agents

The 70 tools

Claude.ai, Claude Desktop and Cowork are offered 67 of these: the three AI image generation tools are withheld there, and in Claude Code unless its connection was registered dynamically. Other clients can get all 70. Read tools change nothing; write tools carry the hints that let clients ask before running them.

Repository layout

pixelesq-skills/├── .claude-plugin/│   ├── plugin.json            Claude Code plugin manifest│   └── marketplace.json       lets the repo install as a marketplace├── .cursor-plugin/plugin.json Cursor plugin manifest├── .mcp.json                  the connector for Claude Code (http transport)├── mcp.json                   the same connector in Cursor's format├── AGENTS.md                  context for agents that read AGENTS.md├── CHANGELOG.md├── agents/                    four specialist agents├── assets/logo.svg├── scripts/generate-section-guide.mjs└── skills/    ├── getting-started/    ├── website-management/    references: section-catalog.json, section-guide.md, site-recipes.md    ├── content-creation/      references: quality-rubric.md    ├── design-theming/        references: font-pairings.md, section-visual-guide.md    ├── seo-optimization/      references: json-ld-recipes.md, social-meta-specs.md    ├── analytics-growth/    └── site-operations/       references: pre-launch-checklist.md, webhook-patterns.md

Keeping it current

• The section guides are generated. Refresh skills/website-management/references/section-catalog.json from the connector's list_section_types (all groups, full detail), then run node scripts/generate-section-guide.mjs.
• The tool map mirrors the connector's README, which is generated from the tool registry.
• Validate before publishing: claude plugin validate .
• The ChatGPT plugin is updated with a ZIP on the OpenAI Plugins page (Upload plugin to make changes): run npm run submission:chatgpt in pixelesq-mcp, then node scripts/build-openai-bundle.mjs here writes dist/pixelesq-chatgpt-<version>.zip (the plugin manifest, mcp.json, the skills, the icon and the listing screenshots).

Links

• Pixelesq
• Connector documentation
• Agent Skills
• Model Context Protocol
• Privacy policy and terms
• Support: [email protected] · Help center

License

MIT